MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel
  1. Home
  2. Meeting Program
  3. All Sessions
  4. All Joint Imaging-Therapy General ePoster Discussion
  5. Deep Learning Based Surface Region-Of-Interest Selection for Motion Monitoring in Left Breast Cancer DIBH Radiotherapy

Deep Learning Based Surface Region-Of-Interest Selection for Motion Monitoring in Left Breast Cancer DIBH Radiotherapy

H Chen1,2 , X Zhen2 , L Zhou2 , M Chen1 , W Lu1 , S Jiang1 , X Gu1*, (1) UT Southwestern Medical Center, Dallas, TX,(2) Southern Medical University, Guangzhou, Guangdong

Presentations

(Sunday, 7/29/2018) 3:00 PM - 3:30 PM

Room: Exhibit Hall | Forum 6

Purpose: To develop and validate a novel convolutional neural network (CNN) and transfer learning based automatic region of interest (ROI) selection scheme for accurate motion monitoring in breast cancer deep inspiration and breath hold (DIBH) radiotherapy.

Methods: The DIBH surface extracted from the CT images were represented by the curvature entropy and surface normal and unfolded onto 2D feature maps. 2D ROIs map and the corresponding surface mesh ROIs were randomly extracted from the 2D feature map and 3D DIBH surface, respectively. To train and evaluate the registration error(RE)predictive model, the ground truth RE of the randomly extracted ROIs were calculated by registering the corresponding ROI meshes extracted from the DIBH surfaces before and after simulated various motions including translation, rotation and deformation. The pre-trained VGG-16 was modified and fine-tuned for the RE prediction. ROIs that could cover the left breast were feed to the fine-tuned VGG-16 to select the quasi-optimal ROI based on the equivalent error (EE). Thirty patients were used for the five-fold cross-validation, and ten patients were for RE prediction model testing. The root mean square error (RMSE) and the mean absolute error (MAE) were employed to evaluate the prediction accuracy quantitatively.

Results: The RMSE/MAE of the RE prediction accuracies in the five-fold validation and testing are <1mm/0.7mm, <0.5°/0.4°，and <2.7mm/1.7mm for translation, rotation and EE, respectively. Compared with the ground truth ROI, which was chosen with the minimal EE, the mean RE difference between the selected ROI via our method and the optimal ROI is <1mm and <0.5° for translation and rotation, respectively.

Conclusion: The satisfactory performances achieved have demonstrated the effectiveness and accuracy of the proposed RE prediction model utilizing information from surface curvature and direction features, indicating it a potential ROI selection tool for left breast cancer DIBH radiotherapy.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email

Additional Links